Liquid AI Publishes GGUF Builds of LFM2.5-24B-A2B, Including a 13.6 GB Quantization-Aware Q4_0 Checkpoint

Liquid AI has published GGUF-format builds of LFM2.5-24B-A2B, a 24-billion-parameter model on the lfm2moe architecture. File sizes range from 13.6 GB (Q4_0) to 48 GB (BF16/F16). The Q4_0 file is a Quantization-Aware Distillation (QAD) checkpoint, which Liquid AI says was trained to recover the accuracy lost to Q4_0 quantization.

Available quantizations

The repository lists the following files:

Format	Size
Q4_0 (QAD checkpoint)	13.6 GB
Q4_K_M	14.5 GB
Q5_K_M	17 GB
Q6_K	19.7 GB
Q8_0	25.5 GB
BF16	48 GB
F16	48 GB

The Q4_0 build is named LFM2.5-24B-A2B-QAD-Q4_0.gguf. Unlike a standard post-hoc Q4_0 conversion, it comes from a distillation step performed with the quantization format in the loop. Liquid AI has not published accuracy figures comparing the QAD build with standard Q4_0 or Q4_K_M in this repository.

Running it

The model card gives llama.cpp usage:

llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF --temp 0.1 --top-k 50 --repeat-penalty 1.05

To select the QAD build explicitly, add --hf-file LFM2.5-24B-A2B-QAD-Q4_0.gguf. These sampling parameters match the base model's generation_config.json. Each GGUF also embeds them as general.sampling.* metadata, so llama.cpp applies them by default.

Model lineage

The repository sits in the following tree:

• Base model: LiquidAI/LFM2.5-24B-A2B-Base
• Fine-tuned: LiquidAI/LFM2.5-24B-A2B
• This repository: GGUF quantizations of the fine-tuned model

Hugging Face lists 62 quantized derivatives and one fine-tune of the model. The repository recorded 45,676 downloads over the last month. It belongs to a collection of post-trained and base LFM2.5 models with 18 items, last updated two days before the page snapshot. No Inference Provider currently hosts it.

Liquid AI describes the LFM2 family as hybrid models designed for edge AI and on-device deployment. It claims they set a new standard for quality, speed and memory efficiency. This page contains no benchmark data to support that claim.

Not disclosed on this page

• Context window
• Active parameter count (the "A2B" suffix suggests roughly 2B active parameters per token, but the page does not state it)
• Benchmark scores
• Training data and training cutoff date
• Licensing terms
• Pricing (the model is downloadable weights and has no hosted API)

What this means

The sizes are the main practical detail. A 24B model at 13.6 GB fits on a single 16 GB-class GPU or a unified-memory machine with room for KV cache, and the Q4_K_M and Q5_K_M builds add little overhead. If the "A2B" suffix does mean about 2B active parameters, per-token compute would be far below that of a dense 24B model. That is the usual argument for mixture-of-experts models on edge hardware, though memory footprint still scales with total parameters.

The QAD Q4_0 checkpoint is the more interesting engineering choice. Most local-inference users rely on community quantizations made after training. A vendor-supplied, distillation-trained 4-bit file could narrow the quality gap, but that depends on published comparisons, and none are included here.

The 45,676 monthly downloads and 62 community quantizations show real uptake. Builders should still wait for independent benchmarks, context-length testing and licensing details before committing to LFM2.5-24B-A2B in production.